Interface for AI to Control Your Computer
Give AI hands on your OS — safely
AGINT is an open-source interface layer that empowers Large Language Models to operate your computer directly: mouse, keyboard, filesystem, browser, apps — with auditable logs, permissions, and human-in-the-loop control.
Five-Minute First Task
Download, run, connect an LLM, and execute your first automated task within minutes — no complex setup.
Model-Agnostic
Use local or cloud LLMs. Swap models without changing your automation code via a thin adapter layer.
Safe by Design
Interactive permission prompts, sandboxed actions, and human-in-the-loop checkpoints keep you in control.
How It Works
- Connect LLM. Provide an API key or use a local runtime.
- Describe a task. Natural language in CLI or UI.
- Planning. The planner breaks the task into safe, auditable actions.
- Execution. Mouse, keyboard, browser, and filesystem tools do the work.
- Review. You approve, pause, or revert with full logs and screenshots.
Core Architecture
┌─ agint/
│ ├─ core/ # event bus, sessions, logging
│ ├─ tools/ # mouse, keyboard, browser, fs, shell
│ ├─ models/ # adapters: OpenAI, Claude, Mistral, local
│ ├─ planner/ # task → actions, safety checks
│ ├─ ui/ # web / desktop UI
│ └─ cli/ # agint & helpers
└──────────────────────────────────────────────Permissions & Policies
Each action requires explicit capability. Scope tokens limit what the agent can do during a session.
Sandbox & Boundaries
Restricted filesystem paths, network allow-lists, and rate-limited UI controls reduce risk.
Full Audit Trail
Deterministic replay, screenshots, and signed logs make every run reviewable and reversible.
Support the Project
Help us build the future of AI-computer interaction. Every contribution brings us closer to a fully functional, secure, and open-source solution.
Crypto Wallet Address:
Investment Opportunities
We are actively seeking investors to accelerate AGINT's development and bring AI-powered computer automation to mainstream adoption.
What We Offer
Investment Use
Plugins
Drop-in actions: open_app, type, click, read_text, select, wait_for, http, shell, git, and more.
Cross-Platform
Targeting Windows, Linux, macOS with a unified action spec and per-OS drivers.
Observability
Prometheus metrics, structured logs, and error traces for reliable automations.
Roadmap
Core & CLI. Mouse, keyboard, fs, browser; session logs; planner v1; demo UI.
Security. Capabilities, sandboxing, network & file scopes; replayer; policy DSL.
Ecosystem. Plugin registry, templates, tutorials, benchmark suite, community examples.
FAQ
Is AGINT safe to run?
Yes — actions require consent and are constrained by policies and scopes. You can dry-run any task.
Do I need a specific model?
No. Adapters make it work with various LLMs, local or cloud.
Can I contribute?
Absolutely. Good first issues include new tool plugins and tests for the planner.
Early Prototype
We have an early working prototype that demonstrates the core concepts of AGINT. Check it out to see the vision in action.